def instantiate_bunet(params, adam):
input_img = Input(params["img_shape"], name='img')
model = mt.get_bunet(input_img,
n_filters=16,
dropout=0.05,
batchnorm=True,
training=False)
'''Compile model'''
adam = adam(lr=params["learning_rate"],
beta_1=0.9,
beta_2=0.999,
epsilon=None,
decay=0.0,
amsgrad=False)
model.compile(optimizer=adam, loss="mse",
metrics=["mse"]) #,percentual_deviance])
model.summary()
'''train and save model'''
save_model_path = os.path.join(evaluation_params["model_path"],
"weights.hdf5")
'''Load models trained weights'''
model.load_weights(save_model_path)
return model
def return_model(params, model_iter, filters):
'''get model'''
input_img = Input(params["img_shape"], name='img')
if model_iter == 1:
model = mt.get_bunet(input_img,
n_filters=4 * filters,
dropout=params["drop_out"],
batchnorm=True,
training=True)
if model_iter == 2:
model = mt.get_shallow_bunet(input_img,
n_filters=4 * filters,
dropout=params["drop_out"],
batchnorm=True,
training=True)
if model_iter == 3:
model = mt.get_very_shallow_bunet(input_img,
n_filters=4 * filters,
dropout=params["drop_out"],
batchnorm=True,
training=True)
return model
def sample_espistemic_loss(params, IDs):
'''
:param IDs: record ids to be evaluated
:param T: number of times to sample the network weights
:return: a nested list with each row being a record and each sublist
a sample
Inspired by paper: https://arxiv.org/pdf/1703.04977.pdf
'''
from skimage.transform import resize
# init nested list
epi_variances = []
epi_record_id = []
# retrieve each sample
# load model with dropout in training mode
input_img = Input(params["img_shape"], name='img')
epistemic_model = mt.get_bunet(input_img,
n_filters=16,
dropout=0.5,
batchnorm=True,
training=True)
'''train and save model'''
save_model_path = os.path.join(evaluation_params["model_path"],
"weights.hdf5")
'''Load models trained weights'''
epistemic_model.load_weights(save_model_path)
print("loaded model with dropout in training mode")
for i in IDs:
im_path = os.path.join(gen_params["fundus_path"], i + ".png")
lbl_path = os.path.join(gen_params["thickness_path"], i + ".npy")
im, lbl = evaluation_load(im_path, lbl_path)
pred = epistemic_model.predict(im.reshape(1, params["img_shape"][0],
params["img_shape"][1], 1),
batch_size=1)
#get only uncertainties for fovea region
prediction = pred[0, :, :, 0].flatten().tolist()
epi_variances.extend(prediction)
epi_record_id.extend([i] * len(prediction))
return (epi_variances, epi_record_id)
def instantiate_bunet(params,adam, training):
'''
:param params: params stating config info
:param opt: an optimizer for the network
:return: model object for prediction
'''
'''get model'''
input_img = Input(params["img_shape"], name='img')
model = mt.get_bunet(input_img, n_filters=16, dropout=0.5, batchnorm=True, training=training)
adam = adam(lr=params["learning_rate"], beta_1=0.9, beta_2=0.999, epsilon=None, decay=0.0, amsgrad=False)
'''Compile model'''
model.compile(optimizer=adam, loss="sparse_categorical_crossentropy", metrics=[iou])
model.summary()
'''train and save model'''
save_model_path = os.path.join(params["save_path"], "weights.hdf5")
model.load_weights(save_model_path)
return model